My understanding is that it's in vogue to use deep learning for complex control problems, and the results are fairly impressive. The idea is to train robotic motion end to end with RL. Not an expert so I don't know the strength and weaknesses versus classical approaches.